Useful Papers - Smoothing
Hopfinger et. al (2000), "A study of analysis parameters that influence the sensitivity of event-related fMRI analyses," NeuroImage 11, 326-333 PDF
Summary: A group including SPM authors tests out different settings for their analysis and looks at how activation Z-scores vary as a function of spatial smoothing, temporal smoothing, different HRFs, and resampled voxel size. The authors divide the effects of spatial smoothing into two points: 1) suppressing high-frequency spatial noise, 2) directly improving p-values by introducing spatial correlation (which affects corrected p-values only).
Bottom line: In the cortex, 6mm smoothing gave the best results - but that was the smallest kernel they used. In subcortical structures, 10mm spatial smoothing gives the best results, but no info about spatial extent of activations is given, so the results are difficult to interpret.
Skudlarski et. al (1999), "ROC analysis of statistical methods used in functional MRI: individual subjects," NeuroImage 9, 311-329 PDF
Summary: We saw this paper way back in week 1, and we'll see it again. Skudlarski et. al use the receiver operating characteristic (ROC) to measure the sensitivity of their analysis, using fake activations in real single-subject data, as a function of various parameters including spatial smoothing. They highlight the relationship of cluster filtering to smoothing, and introduce a couple new smoothing schemes, like multifiltering (combining smoothed activations with unsmoothed).
Bottom line: Smoothing at the right kernel size for the activation works well, but with unknown activation sizes, a kernel size of 1-2 voxels worked best overall (although this includes quite small activations). Multifiltering preserved small activations well. Cluster filtering didn't improve sensitivity above spatial smoothing at all.
LaConte et. al (2003), "The evaluation of preprocessing choices in single-subject BOLD fMRI using NPAIRS performance metrics," NeuroImage 18, 10-27 PDF
Summary: Similar study to those above, but using a different (and much more complicated) performance stat, based on treating all the steps in analysis like parameters to be estimated and getting estimates of reproducibility from each iteration.
Bottom line: Smoothing is good... I think. Honestly, had a tough time making head or tail of the graphs in this one.
Kiebel & Friston (2002), "Anatomically informed basis functions in multisubject studies," Human Brain Mapping 16, 36-46 PDF
Summary: Kiebel & Friston describe a method of smoothing in which smoothing is done only within the cortical sheet - essentially, making the data smoother where you want it, without spreading signal out into areas you're not interested in. They extend this work to multisubject studies and show that it can increase sensitivity relative to standard smoothing approaches.
Bottom line: A nice look at where smoothing might be going and integrating with segmentation, etc.
Zarahn et. al (1997), "Empirical analyses of BOLD fMRI statistics I: Spatially unsmoothed data collected under null-hypothesis conditions," NeuroImage 5, 179-197 PDF
Summary: One of the original empirical papers examining things like true noise distribution in fMRI; the authors looked at things like spatial and temporal coherence in the data and noise profiles from real subjects at rest and from phantom data.
Bottom line: Spatial noise was mostly just noise, but did contain some coherence, which was far greater at lower temporal frequencies - one of the big sources of noise preventing low-temporal-frequency experiments in fMRI from being super effective.
White et. al (2001), "Anatomic and functional variability: the effects of filter size in group fMRI data analysis," NeuroImage 13, 577-588 PDF
Summary: A more focused study looking at smoothing filter sizes, testing out sizes between 0 and 18 mm. A nice look at exactly what happens to your activation clusters as you gradually increase or decrease filter size, including enlargement of activations and merging of apparently separate clusters...
Also see Desmond & Glover (2002) in DesignPapers, which shows that spatial smoothing at 5 mm reduced their within-subject variability estimates substantially.